Obstacle classification and detection for vision based navigation for autonomous driving

With the rising trend in research and development of autonomous vehicles, it is important to keep in mind the cost effectiveness of the system. The cost of high-end sensor technologies being astronomically expensive, the research opportunities are restricted to a select few of high-tech companies and research laboratories such as Google, Tesla, Ford, and the likes of it. Hence our main focus is to develop an autonomous system suitable for academic and research purposes as well. This can be achieved by using available sensors such as the monocular cameras. The existing computer vision techniques along with the deep learning tools like Convolutional Neural Network (CNN) can together be used for developing a robust vision based autonomous driving system. The proposed method uses the SegNet encoder-decoder architecture for pixel-wise semantic segmentation of the video frame followed by an obstacle detection algorithm. The entire algorithm was implemented and tested on a mobile embedded platform of NVIDIA's Jetson TK1.